Objectives
Now is the time to take the concepts from the first six weeks and apply them to a project idea.
2: Jan 31 - Landmark Detection Models
3: Feb 7 - Image Segmentation Models
4: Feb 14 - Training a Neural Network I
4: Feb 21 - Training a Neural Network II
6: Feb 28 - TensorFlow.js and Feature Extraction
Work
- Link to sketch:
- one-sentence description
- summary
Work Desk
The Idea
- of all the ideas I went through.. (list 1)(list 2) I had so many ideas but so little time.. These were the two I thought worth exploring.
- aww face predictor
- user flow: provide cute photos. takes snapshots every 5ms. do it x times → predict their aww face
- need to pull cutesy images from somewhere, but for that I could just download a lot of images
- another potential challenge was not classifying but generating the predicted face.
- vibe checker
- I had the idea, but for computing the distance I got inspired from: https://artsexperiments.withgoogle.com/xdegrees/ (‣)
- user flow: upload 1 photo each of 2 people. machine computes the distance between two and gives them the compatibility score.
- computation: segment image and get the human image → take 2 ppl and do a feature extraction → PCA → compute distance
Process
- I thought the 2nd idea was more doable, and I spent a lot of time what the segmentation model outputs are. With the Deeplab / ml5 BodySegmentation model we can’t really tell that there is person 1 and person 2 in the photo - thus I thought a user flow of uploading separate image inputs is better.
- Also, to distill a person’s image into a number it requires the segmented out image pixel data but what the segmenting models return look sort of like depth maps. I had to get the pixel info of the segmentation maps then peel off pixels from the input image that was segmented as humans.
Next Steps
